Підвищення коефіцієнта ущільненні зображень на основі двовимірних ортогональних перетворень

Abstract

The features of image compression based on two-dimensional orthogonal transformations are considered. It is shown that the increase in the compression ratio can be achieved through the grouping of transformants of discrete orthogonal transformations (Walsh-Hadamard, DCT, and others) and their vector quantization using a self-organizing neural network proposed by Finnish scientist T. Kohonen (Self-Organizing Feature Map – SOFM). Grouping coefficients is an alternative to JPEG zigzag scanning and can not only improve the accuracy of vector quantization but also increase the compression ratio at the lossless compression stage by improving the accuracy of the modeler in adaptive lossless coding schemes, which can increase the compression ratio. Also, after grouping the coefficients, it is easy to organize zonal selection of coefficients, as the coefficients with the same frequencies are localized in certain areas of the image space, which allows them to be excluded from the image synthesis process. Testing has shown that grouping DCT coefficients increases the compression ratio by 1.2-2 times compared to methods without coefficient grouping. Further increase in the image compression ratio is possible with the application of vector quantization of DCT transformants. Studies have shown that the best results are provided by combined quantization – integer quantization for low-frequency transformants and vector quantization using a two-dimensional Kohonen map for high-frequency ones, which allowed increasing the compression ratio of Grayscale images by 30% compared to the JPEG method